Nature Neuroscience

With dense axonal connectivity to every region of cortex, the thalamus plays a central role in the nervous system from sensory processing to cognitive functions. Yet how tissue maturation of the thalamus unfolds during childhood and contributes to typical or atypical development is not clear. Through several large datasets, we provide here a thalamic portrait of fine-scale structural development whose nuclei develop along unique trajectories, some of which diverge from predictions of developmental theory. We find that those thalamic nuclei which show the most protracted development are at the greatest risk for later clinical differences in schizophrenia. The spatial pattern across thalamic nuclei for early psychosis risk is associated with a unique neuroreceptor fingerprint with implications for symptom severity.

Preprints now disseminate a large share of biomedical research before peer review. Because they have not yet passed peer review, some scientists regard preprint claims as unverified or potentially unreliable, yet how much those claims change before publication has so far been quantified only in smaller cohorts, with results that vary by field and topic. Here, we compiled every bioRxiv preprint posted between 2018 and 2025 that we could match by DOI to a peer-reviewed published version, yielding 72,644 preprint-publication pairs. Using a large language model (Claude Sonnet 4.6), we parsed every preprint-publication abstract pair into one primary and two secondary claims, and classified each pair for content change (unchanged, minor, major) and hedging shift (more cautious, more confident, unchanged). On a validation subsample, the model agreed with two independent domain experts about as well as the experts agreed with each other (Cohens kappa 0.63 to 0.66). The primary claim was unchanged in 39.9% of abstracts, minorly revised in 50.0%, and substantially revised in only 10.2%. Hedging shifts were uncommon and asymmetric, with twice as many claims becoming more cautious as more confident (8.4% vs 4.2%). Major revisions were more frequent after long peer review (14.1% in the slowest versus 7.0% in the fastest tertile of review time) and declined over the study period (17.0% in 2019 to 5.7% in 2024). Over the same period, biomedical papers that were never posted as preprints were retracted at roughly twice the rate of those that were. Together, these data show that the move from preprint to peer-reviewed publication leaves the central claims of most biomedical abstracts intact, indicating that preprints are a reliable source of biomedical research.

Chronic pain engages distributed cortical and subcortical circuits, and large-scale intracranial recordings in humans offer a valuable opportunity to characterize its neural signatures. Here, we recorded multi-day stereoelectroencephalography (sEEG) from six participants with refractory chronic neuropathic pain, each implanted with sEEG electrodes spanning dozens of cortical and subcortical structures. Using simultaneous chronic pain ratings, we decoded spontaneous high versus low pain states within individuals (median area under the curve = 0.72; five of six participants performed above chance). Pain-predictive signals were broadly distributed and highly participant-specific. However, mapping the spatial distribution of pain-predictive features revealed preferential representation within canonical macroscale networks: beta-band activity in the default mode network and high-gamma activity in the salience network. These results demonstrate that intracranial recordings can capture distributed, network-organized representations of spontaneous chronic pain states.

Metabolic psychiatry has recently achieved unprecedented clinical rescue in treatment-resistant Schizophrenia (SCZ) utilizing targeted ketogenic interventions. However, the field has operated without a defined genomic anchor, leaving the biophysical mechanism of these therapies largely unexplained. Here, we report the discovery of the definitive metabolic sensor array driving this pathology. By integrating high-resolution topological mapping of SCZ GWAS summary statistics, 3D chromatin conformation (Hi-C), and multi-tissue transcriptomics, we identify massive, non-coding structural variances flanking the HCAR2/HCAR1 tandem locus--the brain's master thermodynamic governor. We demonstrate that while the protein-coding hardware of these receptors remains intact, their shared 3D Topologically Associating Domain (TAD) is fundamentally fractured. This structural collapse drives a perfect transcriptomic double dissociation in the human cortex: the 3' mutational "skyscraper" severely downregulates the HCAR1 lactate emergency brake, while the 5' mutational cluster selectively paralyzes the HCAR2 beta-hydroxybutyrate (BHB) and niacin cooling switch. This dual-flank enhancer failure elegantly provides a definitive genomic etiology for historical SCZ biomarkers, physically explaining both chronic cerebrospinal fluid lactate pooling and the infamous "absent niacin flush." Furthermore, peripheral eQTL mapping reveals profound antagonistic pleiotropy, characterized by a hyper-activation of the HCAR1 lactate shuttle in the testis, explaining the evolutionary conservation of this metabolically catastrophic architecture. Ultimately, we reframe Schizophrenia not as an intrinsic neurological defect, but as an evolutionary "fuel mismatch." The high-performance cognitive architecture of the hominid brain, evolved for ancestral ketogenic environments, experiences a catastrophic thermodynamic crash when deprived of its requisite BHB coolant by modern, high-glycemic diets.

The ability to delay gratification emerges early in life and is linked to long-term health and economic success. Conversely, high impulsivity, marked by a preference for immediate rewards, can be associated with psychiatric disorders. Although processes underlying human delay discounting have been studied at behavioural and macroscopic neural levels, they remain elusive at the single-neuron level. Specifically, it is unknown how human neurons encode extended delays and predict intertemporal choices, and how these processes are impacted by impulsivity. Here, we record single-neuron activity in the human medial temporal lobe (MTL) to explore decision and delay coding. We identify neurons that predict upcoming decisions in the amygdala and hippocampus. Neurons in the entorhinal cortex and hippocampus encode reward delays, with particularly hippocampal population activity coding prospective temporal periods. Importantly, neuronal activity in impulsive individuals shows diminished prospective temporal coding and predicts decisions only shortly before choices are reported. Our findings reveal how distinct MTL regions contribute to intertemporal decisions and provide insight into the neuronal signatures underlying impulsivity.

An estimated 42% of Americans over age 55 will develop dementia, but the molecular understanding of dementia and neurodegenerative disease is constrained because the living human brain cannot be routinely sampled during disease progression. Olfactory sensory neurons provide a clinically accessible neuronal tissue source with developmental, transcriptional, and disease-relevant links to the central nervous system. Here we describe Gateway, a platform that combines device guided olfactory epithelium biopsy, onsite fixation, and 10x Genomics FLEX RNA profiling to generate single-cell transcriptomic data from living patient neurons. We present a 4-million-cell atlas representing 202 human donors, including healthy controls and individuals with neurodegenerative diseases, and release it as an open resource through CELLxGENE. We define the cellular composition of the human olfactory epithelium and show that Gateway captures neuronal functional and compartmental programs and detects more brain-enriched genes than other clinically accessible transcriptomic sample types. In exploratory analyses of Alzheimers Disease and Parkinsons Disease, we identify dys-regulation of pathways and GWAS-implicated genes related to key neurodegenerative mechanisms such as neuroinflammation, endolysosomal biology, proteostasis, and synaptic maintenance. Together, this atlas and clinical workflow establish living patient olfactory neurons as a scalable complementary modality for neuroscience research, target discovery, and biomarker development in neurodegenerative disease.

Adaptive learning requires distinguishing environmental volatility from observation stochasticity, two sources of uncertainty that demand opposite adjustments to the learning rate but inflate experienced variance similarly. Disentangling them is computationally difficult with no tractable closed-form solution. Particle-filter methods are the natural tool for this kind of joint inference, but their stochastic likelihoods and non-differentiable objectives force derivative-free fitting protocols and discourage the individual-difference analyses central to cognitive modeling, where small effect sizes leave little room for additional estimator noise. We introduce the Categorical Bayes Filter (CBF), a deterministic alternative that preserves the conditional structure of recent particle-filter accounts but replaces the stochastic outer layer with a categorical distribution on a quantile grid parameterized through differentiable Beta quantile functions. The procedure performs evidence maximization with an exact, deterministic marginal likelihood that is fully differentiable in the grid parameters. In a volatility-stochasticity task with N = 643 participants, fitted CBF dispersion parameters reveal a cross-over phenotyping pattern between volatility-blind and stochasticity-blind subjects that is not recoverable from particle-filter parameters fit to the same data under a state-of-the-art protocol. The deterministic structure also yields a trial-by-trial ambiguity signal that predicts response times not used in fitting. More broadly, the approach opens individual-level analyses in cognitive modeling and computational psychiatry that stochastic methods have effectively foreclosed.

Advances in high-throughput neural recording technologies enable simultaneous measurement of activity across multiple brain regions in behaving animals, producing datasets of unprecedented scale and richness. Interpreting these data remains challenging due to the bidirectional and temporally offset nature of inter-regional communication, where feedforward and feedback signals are superimposed within neural populations. We introduce BiXformer, a bidirectional cross-attention transformer that disentangles these interactions by decomposing inter-regional communication into causal and acausal streams using directionally masked attention. By enforcing temporal constraints within attention heads, BiXformer recovers low-dimensional, directed latent dynamics and estimates communication delays without relying on linearity or stationarity assumptions. We validate the model on synthetic datasets with known ground-truth delays, demonstrating accurate recovery of both latent structure and inter-regional timing. Applied to simultaneous neural-behavioral recordings and multi-region neural recordings during a movement task, BiXformer reveals interpretable, temporally structured components consistent with the coexistence of sensory feedback and motor-related signals. These results establish BiXformer as a flexible framework for uncovering dynamic, directed communication in complex neural circuits.

How the spatial organization of synapses contributes to stable learning remains a fundamental question in neuroscience. Using the H01 electron microscopy connectome of human temporal cortex, we found that dendritic spines clustered morphologically, whereas synaptic weights followed a center-elevated, surround-suppressed arrangement along dendrites. A regularized Hebbian model formalized this spatial signature, showing that strong synapses lower the probability that neighboring synapses reach high-weight states. Translating this principle into Spatial Synaptic Regularization (SSR) reduced forgetting and stabilized learning across diverse artificial networks and tasks, including continual visual learning, large language-model knowledge editing, and parameter-efficient adaptation of vision-language models, by preserving high-rank, low-overlap representations. These findings identify spatial synaptic organization as an unrecognized dimension for stabilizing learning and show that structural connectomics can yield actionable AI methods.

Prefrontal neurons simultaneously encode multiple task variables. While many studies reported that various groupings of task features could be detected at the population level, the combination of features encoded by individual neurons seemed random. Here, based on unit recordings with Neuropixel probes in behaving rats, we report that far from being random, the representation of information is highly structured. Specifically, the prefrontal network exhibits multiple coding gradients orthogonal to each other in a multidimensional representational space. In this coding structure, neurons have correlated absolute firing rate modulations by different variables, but the polarity of the modulation by one variable is not predictive of that by others. Moreover, this coding structure is manifest in tasks that probe different behavioral processes, ranging from defensive behaviors to sensory discrimination. Last, we find that the same structured representation is apparent in other neocortical regions, including associative and primary sensory areas.

Learning is thought to arise from synaptic modifications embedded in brain-wide circuits 1-3, yet how such circuits coordinate plasticity to support complex behaviour is not known 4,5. Inspired by deep learning, we propose a theory in which pathway-specific cortical feedback drives dendrite-dependent burst plasticity across cortical hierarchies. We show that this mechanism enables online hierarchical credit assignment and learning of complex image recognition and reward-driven tasks. This theory links credit assignment to cell-type-specific control of dendritic excitation-inhibition balance. In doing so, it provides a unified account of cell-type-specific modulation of synaptic plasticity, learning-dependent changes in interneurons, and neuron-specific dendritic error signals. The theory further predicts that interneurons constrain the dimensionality of error-related feedback, offering a functional rationale for cortex-wide gradients in interneuron density. Taken together, these findings indicate that distinct cortical cell types jointly coordinate learning across hierarchical circuits, connecting synaptic plasticity, circuit-level computation, and behaviour.

Compositional generation underlies the systematic and essentially unlimited construction of complex concepts from simpler parts, as is foundational to intelligent behavior, but its underlying neural mechanisms remain unclear. Here we reveal a neural implementation of hierarchical compositional construction of abstract sequences. We demonstrate that in an open-ended setting with very sparse feedback, rats innately utilize hierarchical composition to construct adaptive action sequences that would have been difficult to discover from scratch. Prefrontal neural population representations of these abstract sequences adhere to a low-dimensional format that encodes the orderly progression of elemental units comprising the sequence while converging to a sequence-general endpoint. Higher-level compositions in the hierarchy are systematically related to their lower-level constituent parts, reusing much of the representation, while providing context separation and satisfying format constraints. These neural representations are geometrically identical across animals, pointing to a convergent solution for how knowledge is hierarchically assembled via a compositional mechanism.

Peripheral nerve injury, whether caused by a cut, crush, or excessive stretch, can cause disordered regeneration and formation of a painful neuroma at the injury site. Regrowing axons may end in a swelling, hypersensitive nerve stump, producing pain along the affected nerve distribution that may outweigh sensory or motor deficits. Although neuromas are common, treatment outcomes are inconsistent and the mechanisms that initiate and sustain neuroma-associated pain remain poorly defined. Many surgical and non-surgical approaches (excision with transposition, capping, sclerosis, cryoablation) are used, yet comparative studies have not identified a superior technique, highlighting biological heterogeneity and knowledge gaps. A proximal nerve crush (axonotmetic injury) performed upstream of a neuroma or transection is proposed to reshape axonal growth and interrupt retrograde injury signaling, potentially shifting the system toward a more regenerative, pain-resolving state. However, how proximal crush remodels long-term programs in the dorsal root ganglion (DRG), and how these changes relate to neuroma-like outcomes, are unclear. Here, we build a single-cell atlas of DRG remodeling after neuroma-forming injury and compare it with two interventions- proximal nerve crush and nerve resection, using scRNA-seq. By resolving transcriptional responses across sensory neurons, satellite glial cells, Schwann cells, and macrophages, we identify intervention-specific glial and immune programs that distinguish permissive regeneration from persistent, pain-associated states and nominate pathways for mechanism-guided neuroma therapies.

In reinforcement learning, an agent learns to map representations of the environment state to predictions of future reward. Most prior work in neuroscience has assumed a fixed representation and studied how reward prediction errors (thought to be conveyed by phasic dopamine signals) are used to update the mapping from representations to predictions. However, work in machine learning has demonstrated that much more powerful predictive systems can be learned by using the errors to update the representations themselves. We study whether the brain does something similar by leveraging simultaneous recordings of striatal projection neurons in the olfactory tubercle (putatively representing state features) and dopamine neurons in the ventral tegmental area. We show that trial-by-trial changes in striatal activity are more consistent with dopamine-driven representation learning than a variety of alternative updating schemes. This result suggests a convergence of representation learning principles in biological and artificial systems.

Learning interpretable low-dimensional representations of whole-brain neuronal dynamics remains a major computational challenge in systems neuroscience. We present a wiring-agnostic deep-learning framework that couples a convolutional encoder with a temporal transformer to learn compact representations directly from volumetric calcium imaging of the entire Drosophila melanogaster brain. Trained to classify 16 experimental conditions that factorially combine metabolic state (fed, starved), sensory modality (olfaction, gustation, or combined), and stimulus valence (appetitive, aversive, or conflicting), the model organizes pan-neuronal whole-brain population activity into geometrically distinct, condition-specific clusters. Analysis of the models latent space reveals that state, modality, and valence are encoded along three near-orthogonal axes: a separable structure that emerges from the classification objective without explicit disentanglement constraints. Spatial attribution and regional importance analyses link modality decoding to distinct anatomical circuits, whereas metabolic state and valence related information show weaker regional specificity and broader distribution across the brain. Our approach does not require anatomical annotation, neuronal identification, or connectivity information, and thus provides a scalable foundation for comparative whole-brain imaging and representation learning of brain wide dynamics.

New implantable-electrode fabrication strategies make dense, ultrafine electrode arrays with lower tissue burden increasingly feasible, shifting a key bottleneck for scalable brain-computer interfaces from electrode placement to readout capacity. Electrode pooling, in which multiple electrodes share a readout channel, could relax this bottleneck by combining extracellular signals before acquisition, but it has remained unclear whether such compression preserves the neural population structure needed for behavioral decoding. Here we evaluate this question using software-emulated electrode pooling in mouse sensorimotor cortex during a cue-guided reach-and-grasp task using a high-density microwire array coupled to a CMOS microelectrode-array platform. Pooled recordings retain forelimb kinematic information more effectively than a channel-matched control that discards electrodes. Pooling reduces the separability of electrode-specific spikes and sorted units, indicating a loss of some neuronal detail, but the mixed signals still preserve task-aligned low-dimensional latent dynamics that support decoding. When readout capacity is fixed, this trade-off allows broader electrode coverage to contribute to behaviorally informative population sampling. Together, these results define electrode pooling as a design trade-off for scalable readout, in which some electrode-specific neuronal information is lost but the population dynamics needed for movement decoding remain accessible.

When and how does insight emerge? We conceptualize insight as a sudden realization arising from restructuring a world model--an internal interpretation linking actions to outcomes. However, this process remains inaccessible even with verbal report. Here we developed inside insight dynamics (IID), a machine-learning framework estimating latent world-model dynamics from behavioral data. We analyzed mouse data from two tasks differing in difficulty and requiring animals to shift from an initial world model to a new one. IID decoded timing of insight-like shifts and evolving reward beliefs within competing world models. We examined how these shifts were acquired through learning. We found that the harder task was better explained by gated learning, in which a new model becomes learnable only after being recognized, whereas the simpler task favored parallel learning, in which candidate models are learned in advance. Thus, IID opens a route to quantifying latent insight dynamics.

Dense electron-microscopy connectomes provide synaptic-resolution maps of neuronal structure and wiring, but learning scalable representations that integrate structure and connectivity for connectome discovery with minimal human intervention remains difficult. Here we present a self-supervised framework for structure-connectivity representation learning in dense connectomes. A hierarchical graph neural network with skeleton decomposition enables contrastive learning from finely sampled FlyWire neuronal skeletons, showing that fine skeletons preserve substantially richer identity information than coarse representations. Coordinate-free topology reduces developmental and geometric confounds, improving clustering and label-efficient inference. We then use learned structural embeddings as continuous descriptors of synaptic partners to construct structure-driven connectivity representations, improving subtype discrimination without predefined partner-type labels. Iterative multi-hop learning further reveals higher-order organization, including hemispheric connectivity lateralization and connectivity-defined subgroups. Attention analysis links these differences to specific synaptic partners. Together, these results establish a self-supervised and scalable framework for discovering neuronal identity and connectome organization in a large-scale dense connectome.

A single neural circuit can exhibit qualitatively different dynamics across trials: one circuit, many flow fields. Standard models treat this trial-to-trial variability as noise around a fixed dynamical system or as discrete switches between regimes, yet neither captures how continuous internal-state variables, such as arousal or engagement, can gradually deform the circuits flow field. We propose that singletrial fitting can be reframed as inferring the low-dimensional control parameters that reshape a shared circuits flow field. We realize this with a low-rank recurrent network in which trial-specific static input biases act as bifurcation parameters: constant within a trial, they deform the flow field without directly driving activity over time. In a teacher-student setting, the model recovers the underlying dynamical system and its bifurcation structure from activity alone. Applied to large-scale recordings of mouse motor cortex during a delayed movement task, the model identifies a disengagement axis that separates engaged from disengaged trials and, when perturbed in silico, causally shifts the flow field between engaged and disengaged regimes. A generative extension reproduces the distribution of single-trial activity, and the inferred latent structure partially transfers across sessions and animals, suggesting shared low-dimensional structure across motor-cortical circuits. Together, these results reframe a methodological problem of fitting single-trial activity as a scientific opportunity: reading off the control parameters of the underlying dynamics, and connecting data-driven inference of neural dynamics to mechanistic theories of how a single circuit reuses its dynamics for flexible behavior.

Sporadic amyotrophic lateral sclerosis (sALS) lacks longitudinal molecular measurements, making it difficult to distinguish early disease-organising changes from downstream consequences. We present a training-free framework that extracts directional structure from static single-nucleus RNA-seq by applying discrete Hodge decomposition to gene co-expression dynamics across pseudotime-ordered donor states. The framework separates irreversible co-expression cascades from circular feedback structure and regresses out the component explained by the healthy co-expression network, allowing disease-specific organisation to be examined in isolation. Perturbation benchmarks show that experimentally imposed sources are recoverable from control-normalised off-diagonal covariance-response fields, whereas marginal variance and diagonal covariance controls do not recover the source. Applied to sALS primary motor cortex (24 donors, 10 cell types), the framework identifies oligodendrocytes as the most structurally upstream cell type and upper-motor-neuron-containing layers as the most structurally downstream (Oligo cell-type{varphi} = 0.900, with glial cell types preserving the healthy co-expression network topology, whereas neuronal cell types show collapse-dominant deformation). Cytoplasmic translation is the only pathway with reproducible cross-cell-type upstream enrichment. At the gene level, the ribosome-associated quality-control factor NEMF -- which appends C-terminal alanine-threonine tags ("CATylation") to nascent chains on stalled ribosomes -- shows disease-specific loss of co-expression coherence in seven of ten cell types despite essentially unchanged mRNA expression; the disease signal is decoupling from collision-response partners (GCN2, PKR), not expression-level change. Cross-cohort validation across three BA4 motor cortex cohorts (including two external cohorts; total N=107) reproduced the oligodendrocyte-upstream / upper-motor-neuron-downstream structural architecture (Oligo-preserved / ET-sink) in all three cohorts, with NEMF co-expression coherence loss replicated in two of three cohorts. These data support a brain-side, circuit-distal structural model in which oligodendrocyte-lineage stress occupies an upstream-like preserved compartment, while upper-motor-neuron-containing excitatory populations form a downstream sink. The pattern is consistent with -- but does not directly establish -- a cascade architecture in which oligodendrocyte stress structurally precedes motor neuron TDP-43 pathology, and would produce a clinical phenotype resembling dying-back (the conventional view of ALS, in which motor neuron pathology appears to begin at distal axons and spread retrogradely toward the cell body) yet originating centrally and glially. NEMF/CATylation network disruption is identified as a candidate intermediate structural node bridging oligodendrocyte stress and motor neuron TDP-43 pathology.